🚀 go-pugleaf

             DEPARTMENT OF COMPUTER SCIENCE
               COLORADO STATE UNIVERSITY

                       BARNEY'S
             MONDAY AFTERNOON CLUB SEMINAR

	Speaker:	Lucas Roh
			Computer Scientist
			Argonne National Laboratory

	   Date:	November 4, 1996
	   Time:	3:30 - 4:30 p.m.
	  Place:	100 USC

	   Refreshments:  following the talk

  AN EXTENSIBLE AUTOMATIC DIFFERENTIATION TOOL FOR C

   In computational science and engineering, we often 
require the computation of derivatives of a function or 
process.  Automatic differentiation (AD) is a technique 
to evaluate derivatives of the function written as a 
computer algorithm by applying the chain rule to ele-
mentary operations.  In contrast to the divided differ-
ences, AD is accurate; unlike the symbolic differentia-
tion, AD works for a general algorithm that contains 
arbitrary loops and conditionals.

   In this talk, I will describe our ADIC (Automatic 
Differentiation in C) tool.  ADIC takes a C code and 
produces a new C code that computes the original results 
along with the derivatives of the specified input vari-
ables.  Unlike translators, ADIC augments the original 
semantic of the program with additional semantics.  Some 
of the major challenges in dealing with various "warts" 
of the C language, such as side-effects and aliasing, 
will also be discussed.

   AD is still in its infancy.  More sophisticated 
algorithms that apply the chain rule can significantly 
boost the performance of the derivative code.  ADIC 
incorporates a component architecture where these algo-
rithms are embedded in external modules, thus enabling 
experimentation with different algorithms and also 
allowing the sharing of components across tools.  I will 
discuss the component architecture along with modules 
that we have developed including a Hessian module that 
generates the second order derivatives.